An Agentic AI Scientific Community for Automated Neural Operator Discovery

This paper introduces an agentic approach to autonomous neural operator discovery using an AI scientific community framework. The system consists of multiple virtual laboratories that interact under a citation-based economy of influence, where highly-cited labs found new labs following their research direction while replacing underperforming ones. Each virtual lab contains three specialized agents: an LLM planner that proposes neural operator architectures, a numerical worker that trains and evaluates them, and an LLM reviewer that provides peer-review style assessments. This multi-agent architecture enables automated exploration and optimization of neural operator designs for solving partial differential equations in scientific computing.